Few-shot class-incremental learning often assumes homogeneous domains and balanced data, unlike real-world streams. We identify Domain Gravity, where dominant domains distort shared representations, and introduce XD-VSCIL to study this setting. We propose HyCal, a training-free calibration method combining cosine similarity and Mahalanobis distance on frozen CLIP features. HyCal improves retention and adaptation across heterogeneous, imbalanced domains without model updates.

Eunju Lee received a B.S. and M.S. degree in imaging engineering from Chung-Ang University, Seoul, Korea, in 2020 and 2022. She is currently a Ph.D. student in imaging engineering at the Graduate School of Advanced Imaging Science, Multimedia and Film at Chung-Ang University. Her current research focuses on deep learning and computer vision.